Relational attention network for few-shot fault diagnosis of rotating machinery

In practical rotating machinery fault diagnosis (FD) tasks, the scarcity of labeled fault data often impedes the model’s ability to learn effective classification boundaries, resulting in poor performance with limited training samples. To address this issue, this paper introduces a rotation machinery few-shot FD algorithm based on a Relational Attention Network (RAN). The network comprises four modules: feature embedding, feature concatenation, attention computation, and relation computation. RAN achieves a large-margin feature space by utilizing extended concatenation and attention on feature pairs. Specifically, the feature embedding module extracts signal features and constructs both a prototype-query feature pair matrix and a query-query feature pair matrix. Subsequently, the attention computation module assigns higher weights to indistinguishable heterogeneous feature pairs within the two matrices. Finally, the attention-weighted feature pair matrices are fed into the relation computation module to generate a relation score matrix. The effectiveness and superiority of this method were validated on a bearing fault dataset. Experimental results demonstrate that RAN outperforms various comparison methods in recognition accuracy under limited training sample conditions.